mm/sparse.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * sparse memory mappings.
   4  */
   5 #include <linux/mm.h>
   6 #include <linux/slab.h>
   7 #include <linux/mmzone.h>
   8 #include <linux/memblock.h>
   9 #include <linux/compiler.h>
  10 #include <linux/highmem.h>
  11 #include <linux/export.h>
  12 #include <linux/spinlock.h>
  13 #include <linux/vmalloc.h>
  14 #include <linux/swap.h>
  15 #include <linux/swapops.h>
  16 #include <linux/bootmem_info.h>
  17
  18 #include "internal.h"
  19 #include <asm/dma.h>
  20
  21 /*
  22  * Permanent SPARSEMEM data:
  23  *
  24  * 1) mem_section       - memory sections, mem_map's for valid memory
  25  */
  26 #ifdef CONFIG_SPARSEMEM_EXTREME
  27 struct mem_section **mem_section;
  28 #else
  29 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  30         ____cacheline_internodealigned_in_smp;
  31 #endif
  32 EXPORT_SYMBOL(mem_section);
  33
  34 #ifdef NODE_NOT_IN_PAGE_FLAGS
  35 /*
  36  * If we did not store the node number in the page then we have to
  37  * do a lookup in the section_to_node_table in order to find which
  38  * node the page belongs to.
  39  */
  40 #if MAX_NUMNODES <= 256
  41 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  42 #else
  43 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  44 #endif
  45
  46 int page_to_nid(const struct page *page)
  47 {
  48         return section_to_node_table[page_to_section(page)];
  49 }
  50 EXPORT_SYMBOL(page_to_nid);
  51
  52 static void set_section_nid(unsigned long section_nr, int nid)
  53 {
  54         section_to_node_table[section_nr] = nid;
  55 }
  56 #else /* !NODE_NOT_IN_PAGE_FLAGS */
  57 static inline void set_section_nid(unsigned long section_nr, int nid)
  58 {
  59 }
  60 #endif
  61
  62 #ifdef CONFIG_SPARSEMEM_EXTREME
  63 static noinline struct mem_section __ref *sparse_index_alloc(int nid)
  64 {
  65         struct mem_section *section = NULL;
  66         unsigned long array_size = SECTIONS_PER_ROOT *
  67                                    sizeof(struct mem_section);
  68
  69         if (slab_is_available()) {
  70                 section = kzalloc_node(array_size, GFP_KERNEL, nid);
  71         } else {
  72                 section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
  73                                               nid);
  74                 if (!section)
  75                         panic("%s: Failed to allocate %lu bytes nid=%d\n",
  76                               __func__, array_size, nid);
  77         }
  78
  79         return section;
  80 }
  81
  82 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  83 {
  84         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  85         struct mem_section *section;
  86
  87         /*
  88          * An existing section is possible in the sub-section hotplug
  89          * case. First hot-add instantiates, follow-on hot-add reuses
  90          * the existing section.
  91          *
  92          * The mem_hotplug_lock resolves the apparent race below.
  93          */
  94         if (mem_section[root])
  95                 return 0;
  96
  97         section = sparse_index_alloc(nid);
  98         if (!section)
  99                 return -ENOMEM;
 100
 101         mem_section[root] = section;
 102
 103         return 0;
 104 }
 105 #else /* !SPARSEMEM_EXTREME */
 106 static inline int sparse_index_init(unsigned long section_nr, int nid)
 107 {
 108         return 0;
 109 }
 110 #endif
 111
 112 /*
 113  * During early boot, before section_mem_map is used for an actual
 114  * mem_map, we use section_mem_map to store the section's NUMA
 115  * node.  This keeps us from having to use another data structure.  The
 116  * node information is cleared just before we store the real mem_map.
 117  */
 118 static inline unsigned long sparse_encode_early_nid(int nid)
 119 {
 120         return ((unsigned long)nid << SECTION_NID_SHIFT);
 121 }
 122
 123 static inline int sparse_early_nid(struct mem_section *section)
 124 {
 125         return (section->section_mem_map >> SECTION_NID_SHIFT);
 126 }
 127
 128 /* Validate the physical addressing limitations of the model */
 129 static void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 130                                                 unsigned long *end_pfn)
 131 {
 132         unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 133
 134         /*
 135          * Sanity checks - do not allow an architecture to pass
 136          * in larger pfns than the maximum scope of sparsemem:
 137          */
 138         if (*start_pfn > max_sparsemem_pfn) {
 139                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 140                         "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 141                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 142                 WARN_ON_ONCE(1);
 143                 *start_pfn = max_sparsemem_pfn;
 144                 *end_pfn = max_sparsemem_pfn;
 145         } else if (*end_pfn > max_sparsemem_pfn) {
 146                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 147                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 148                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 149                 WARN_ON_ONCE(1);
 150                 *end_pfn = max_sparsemem_pfn;
 151         }
 152 }
 153
 154 /*
 155  * There are a number of times that we loop over NR_MEM_SECTIONS,
 156  * looking for section_present() on each.  But, when we have very
 157  * large physical address spaces, NR_MEM_SECTIONS can also be
 158  * very large which makes the loops quite long.
 159  *
 160  * Keeping track of this gives us an easy way to break out of
 161  * those loops early.
 162  */
 163 unsigned long __highest_present_section_nr;
 164 static void __section_mark_present(struct mem_section *ms,
 165                 unsigned long section_nr)
 166 {
 167         if (section_nr > __highest_present_section_nr)
 168                 __highest_present_section_nr = section_nr;
 169
 170         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 171 }
 172
 173 #define for_each_present_section_nr(start, section_nr)          \
 174         for (section_nr = next_present_section_nr(start-1);     \
 175              section_nr != -1;                                                          \
 176              section_nr = next_present_section_nr(section_nr))
 177
 178 static inline unsigned long first_present_section_nr(void)
 179 {
 180         return next_present_section_nr(-1);
 181 }
 182
 183 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 184 static void subsection_mask_set(unsigned long *map, unsigned long pfn,
 185                 unsigned long nr_pages)
 186 {
 187         int idx = subsection_map_index(pfn);
 188         int end = subsection_map_index(pfn + nr_pages - 1);
 189
 190         bitmap_set(map, idx, end - idx + 1);
 191 }
 192
 193 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
 194 {
 195         int end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
 196         unsigned long nr, start_sec = pfn_to_section_nr(pfn);
 197
 198         if (!nr_pages)
 199                 return;
 200
 201         for (nr = start_sec; nr <= end_sec; nr++) {
 202                 struct mem_section *ms;
 203                 unsigned long pfns;
 204
 205                 pfns = min(nr_pages, PAGES_PER_SECTION
 206                                 - (pfn & ~PAGE_SECTION_MASK));
 207                 ms = __nr_to_section(nr);
 208                 subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
 209
 210                 pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
 211                                 pfns, subsection_map_index(pfn),
 212                                 subsection_map_index(pfn + pfns - 1));
 213
 214                 pfn += pfns;
 215                 nr_pages -= pfns;
 216         }
 217 }
 218 #else
 219 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
 220 {
 221 }
 222 #endif
 223
 224 /* Record a memory area against a node. */
 225 static void __init memory_present(int nid, unsigned long start, unsigned long end)
 226 {
 227         unsigned long pfn;
 228
 229 #ifdef CONFIG_SPARSEMEM_EXTREME
 230         if (unlikely(!mem_section)) {
 231                 unsigned long size, align;
 232
 233                 size = sizeof(struct mem_section *) * NR_SECTION_ROOTS;
 234                 align = 1 << (INTERNODE_CACHE_SHIFT);
 235                 mem_section = memblock_alloc(size, align);
 236                 if (!mem_section)
 237                         panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
 238                               __func__, size, align);
 239         }
 240 #endif
 241
 242         start &= PAGE_SECTION_MASK;
 243         mminit_validate_memmodel_limits(&start, &end);
 244         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 245                 unsigned long section = pfn_to_section_nr(pfn);
 246                 struct mem_section *ms;
 247
 248                 sparse_index_init(section, nid);
 249                 set_section_nid(section, nid);
 250
 251                 ms = __nr_to_section(section);
 252                 if (!ms->section_mem_map) {
 253                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 254                                                         SECTION_IS_ONLINE;
 255                         __section_mark_present(ms, section);
 256                 }
 257         }
 258 }
 259
 260 /*
 261  * Mark all memblocks as present using memory_present().
 262  * This is a convenience function that is useful to mark all of the systems
 263  * memory as present during initialization.
 264  */
 265 static void __init memblocks_present(void)
 266 {
 267         unsigned long start, end;
 268         int i, nid;
 269
 270         for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid)
 271                 memory_present(nid, start, end);
 272 }
 273
 274 /*
 275  * Subtle, we encode the real pfn into the mem_map such that
 276  * the identity pfn - section_mem_map will return the actual
 277  * physical page frame number.
 278  */
 279 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 280 {
 281         unsigned long coded_mem_map =
 282                 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 283         BUILD_BUG_ON(SECTION_MAP_LAST_BIT > PFN_SECTION_SHIFT);
 284         BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
 285         return coded_mem_map;
 286 }
 287
 288 #ifdef CONFIG_MEMORY_HOTPLUG
 289 /*
 290  * Decode mem_map from the coded memmap
 291  */
 292 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 293 {
 294         /* mask off the extra low bits of information */
 295         coded_mem_map &= SECTION_MAP_MASK;
 296         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 297 }
 298 #endif /* CONFIG_MEMORY_HOTPLUG */
 299
 300 static void __meminit sparse_init_one_section(struct mem_section *ms,
 301                 unsigned long pnum, struct page *mem_map,
 302                 struct mem_section_usage *usage, unsigned long flags)
 303 {
 304         ms->section_mem_map &= ~SECTION_MAP_MASK;
 305         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
 306                 | SECTION_HAS_MEM_MAP | flags;
 307         ms->usage = usage;
 308 }
 309
 310 static unsigned long usemap_size(void)
 311 {
 312         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 313 }
 314
 315 size_t mem_section_usage_size(void)
 316 {
 317         return sizeof(struct mem_section_usage) + usemap_size();
 318 }
 319
 320 #ifdef CONFIG_MEMORY_HOTREMOVE
 321 static inline phys_addr_t pgdat_to_phys(struct pglist_data *pgdat)
 322 {
 323 #ifndef CONFIG_NUMA
 324         VM_BUG_ON(pgdat != &contig_page_data);
 325         return __pa_symbol(&contig_page_data);
 326 #else
 327         return __pa(pgdat);
 328 #endif
 329 }
 330
 331 static struct mem_section_usage * __init
 332 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 333                                          unsigned long size)
 334 {
 335         struct mem_section_usage *usage;
 336         unsigned long goal, limit;
 337         int nid;
 338         /*
 339          * A page may contain usemaps for other sections preventing the
 340          * page being freed and making a section unremovable while
 341          * other sections referencing the usemap remain active. Similarly,
 342          * a pgdat can prevent a section being removed. If section A
 343          * contains a pgdat and section B contains the usemap, both
 344          * sections become inter-dependent. This allocates usemaps
 345          * from the same section as the pgdat where possible to avoid
 346          * this problem.
 347          */
 348         goal = pgdat_to_phys(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 349         limit = goal + (1UL << PA_SECTION_SHIFT);
 350         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 351 again:
 352         usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
 353         if (!usage && limit) {
 354                 limit = 0;
 355                 goto again;
 356         }
 357         return usage;
 358 }
 359
 360 static void __init check_usemap_section_nr(int nid,
 361                 struct mem_section_usage *usage)
 362 {
 363         unsigned long usemap_snr, pgdat_snr;
 364         static unsigned long old_usemap_snr;
 365         static unsigned long old_pgdat_snr;
 366         struct pglist_data *pgdat = NODE_DATA(nid);
 367         int usemap_nid;
 368
 369         /* First call */
 370         if (!old_usemap_snr) {
 371                 old_usemap_snr = NR_MEM_SECTIONS;
 372                 old_pgdat_snr = NR_MEM_SECTIONS;
 373         }
 374
 375         usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
 376         pgdat_snr = pfn_to_section_nr(pgdat_to_phys(pgdat) >> PAGE_SHIFT);
 377         if (usemap_snr == pgdat_snr)
 378                 return;
 379
 380         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 381                 /* skip redundant message */
 382                 return;
 383
 384         old_usemap_snr = usemap_snr;
 385         old_pgdat_snr = pgdat_snr;
 386
 387         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 388         if (usemap_nid != nid) {
 389                 pr_info("node %d must be removed before remove section %ld\n",
 390                         nid, usemap_snr);
 391                 return;
 392         }
 393         /*
 394          * There is a circular dependency.
 395          * Some platforms allow un-removable section because they will just
 396          * gather other removable sections for dynamic partitioning.
 397          * Just notify un-removable section's number here.
 398          */
 399         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 400                 usemap_snr, pgdat_snr, nid);
 401 }
 402 #else
 403 static struct mem_section_usage * __init
 404 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 405                                          unsigned long size)
 406 {
 407         return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
 408 }
 409
 410 static void __init check_usemap_section_nr(int nid,
 411                 struct mem_section_usage *usage)
 412 {
 413 }
 414 #endif /* CONFIG_MEMORY_HOTREMOVE */
 415
 416 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 417 static unsigned long __init section_map_size(void)
 418 {
 419         return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
 420 }
 421
 422 #else
 423 static unsigned long __init section_map_size(void)
 424 {
 425         return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 426 }
 427
 428 struct page __init *__populate_section_memmap(unsigned long pfn,
 429                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
 430                 struct dev_pagemap *pgmap)
 431 {
 432         unsigned long size = section_map_size();
 433         struct page *map = sparse_buffer_alloc(size);
 434         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 435
 436         if (map)
 437                 return map;
 438
 439         map = memmap_alloc(size, size, addr, nid, false);
 440         if (!map)
 441                 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
 442                       __func__, size, PAGE_SIZE, nid, &addr);
 443
 444         return map;
 445 }
 446 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 447
 448 static void *sparsemap_buf __meminitdata;
 449 static void *sparsemap_buf_end __meminitdata;
 450
 451 static inline void __meminit sparse_buffer_free(unsigned long size)
 452 {
 453         WARN_ON(!sparsemap_buf || size == 0);
 454         memblock_free(sparsemap_buf, size);
 455 }
 456
 457 static void __init sparse_buffer_init(unsigned long size, int nid)
 458 {
 459         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 460         WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
 461         /*
 462          * Pre-allocated buffer is mainly used by __populate_section_memmap
 463          * and we want it to be properly aligned to the section size - this is
 464          * especially the case for VMEMMAP which maps memmap to PMDs
 465          */
 466         sparsemap_buf = memmap_alloc(size, section_map_size(), addr, nid, true);
 467         sparsemap_buf_end = sparsemap_buf + size;
 468 }
 469
 470 static void __init sparse_buffer_fini(void)
 471 {
 472         unsigned long size = sparsemap_buf_end - sparsemap_buf;
 473
 474         if (sparsemap_buf && size > 0)
 475                 sparse_buffer_free(size);
 476         sparsemap_buf = NULL;
 477 }
 478
 479 void * __meminit sparse_buffer_alloc(unsigned long size)
 480 {
 481         void *ptr = NULL;
 482
 483         if (sparsemap_buf) {
 484                 ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
 485                 if (ptr + size > sparsemap_buf_end)
 486                         ptr = NULL;
 487                 else {
 488                         /* Free redundant aligned space */
 489                         if ((unsigned long)(ptr - sparsemap_buf) > 0)
 490                                 sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
 491                         sparsemap_buf = ptr + size;
 492                 }
 493         }
 494         return ptr;
 495 }
 496
 497 void __weak __meminit vmemmap_populate_print_last(void)
 498 {
 499 }
 500
 501 /*
 502  * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
 503  * And number of present sections in this node is map_count.
 504  */
 505 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
 506                                    unsigned long pnum_end,
 507                                    unsigned long map_count)
 508 {
 509         struct mem_section_usage *usage;
 510         unsigned long pnum;
 511         struct page *map;
 512
 513         usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
 514                         mem_section_usage_size() * map_count);
 515         if (!usage) {
 516                 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
 517                 goto failed;
 518         }
 519         sparse_buffer_init(map_count * section_map_size(), nid);
 520         for_each_present_section_nr(pnum_begin, pnum) {
 521                 unsigned long pfn = section_nr_to_pfn(pnum);
 522
 523                 if (pnum >= pnum_end)
 524                         break;
 525
 526                 map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
 527                                 nid, NULL, NULL);
 528                 if (!map) {
 529                         pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
 530                                __func__, nid);
 531                         pnum_begin = pnum;
 532                         sparse_buffer_fini();
 533                         goto failed;
 534                 }
 535                 check_usemap_section_nr(nid, usage);
 536                 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
 537                                 SECTION_IS_EARLY);
 538                 usage = (void *) usage + mem_section_usage_size();
 539         }
 540         sparse_buffer_fini();
 541         return;
 542 failed:
 543         /* We failed to allocate, mark all the following pnums as not present */
 544         for_each_present_section_nr(pnum_begin, pnum) {
 545                 struct mem_section *ms;
 546
 547                 if (pnum >= pnum_end)
 548                         break;
 549                 ms = __nr_to_section(pnum);
 550                 ms->section_mem_map = 0;
 551         }
 552 }
 553
 554 /*
 555  * Allocate the accumulated non-linear sections, allocate a mem_map
 556  * for each and record the physical to section mapping.
 557  */
 558 void __init sparse_init(void)
 559 {
 560         unsigned long pnum_end, pnum_begin, map_count = 1;
 561         int nid_begin;
 562
 563         memblocks_present();
 564
 565         pnum_begin = first_present_section_nr();
 566         nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
 567
 568         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 569         set_pageblock_order();
 570
 571         for_each_present_section_nr(pnum_begin + 1, pnum_end) {
 572                 int nid = sparse_early_nid(__nr_to_section(pnum_end));
 573
 574                 if (nid == nid_begin) {
 575                         map_count++;
 576                         continue;
 577                 }
 578                 /* Init node with sections in range [pnum_begin, pnum_end) */
 579                 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 580                 nid_begin = nid;
 581                 pnum_begin = pnum_end;
 582                 map_count = 1;
 583         }
 584         /* cover the last node */
 585         sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 586         vmemmap_populate_print_last();
 587 }
 588
 589 #ifdef CONFIG_MEMORY_HOTPLUG
 590
 591 /* Mark all memory sections within the pfn range as online */
 592 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 593 {
 594         unsigned long pfn;
 595
 596         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 597                 unsigned long section_nr = pfn_to_section_nr(pfn);
 598                 struct mem_section *ms;
 599
 600                 /* onlining code should never touch invalid ranges */
 601                 if (WARN_ON(!valid_section_nr(section_nr)))
 602                         continue;
 603
 604                 ms = __nr_to_section(section_nr);
 605                 ms->section_mem_map |= SECTION_IS_ONLINE;
 606         }
 607 }
 608
 609 /* Mark all memory sections within the pfn range as offline */
 610 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 611 {
 612         unsigned long pfn;
 613
 614         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 615                 unsigned long section_nr = pfn_to_section_nr(pfn);
 616                 struct mem_section *ms;
 617
 618                 /*
 619                  * TODO this needs some double checking. Offlining code makes
 620                  * sure to check pfn_valid but those checks might be just bogus
 621                  */
 622                 if (WARN_ON(!valid_section_nr(section_nr)))
 623                         continue;
 624
 625                 ms = __nr_to_section(section_nr);
 626                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 627         }
 628 }
 629
 630 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 631 static struct page * __meminit populate_section_memmap(unsigned long pfn,
 632                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
 633                 struct dev_pagemap *pgmap)
 634 {
 635         return __populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
 636 }
 637
 638 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
 639                 struct vmem_altmap *altmap)
 640 {
 641         unsigned long start = (unsigned long) pfn_to_page(pfn);
 642         unsigned long end = start + nr_pages * sizeof(struct page);
 643
 644         vmemmap_free(start, end, altmap);
 645 }
 646 static void free_map_bootmem(struct page *memmap)
 647 {
 648         unsigned long start = (unsigned long)memmap;
 649         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 650
 651         vmemmap_free(start, end, NULL);
 652 }
 653
 654 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
 655 {
 656         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
 657         DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
 658         struct mem_section *ms = __pfn_to_section(pfn);
 659         unsigned long *subsection_map = ms->usage
 660                 ? &ms->usage->subsection_map[0] : NULL;
 661
 662         subsection_mask_set(map, pfn, nr_pages);
 663         if (subsection_map)
 664                 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
 665
 666         if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
 667                                 "section already deactivated (%#lx + %ld)\n",
 668                                 pfn, nr_pages))
 669                 return -EINVAL;
 670
 671         bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
 672         return 0;
 673 }
 674
 675 static bool is_subsection_map_empty(struct mem_section *ms)
 676 {
 677         return bitmap_empty(&ms->usage->subsection_map[0],
 678                             SUBSECTIONS_PER_SECTION);
 679 }
 680
 681 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
 682 {
 683         struct mem_section *ms = __pfn_to_section(pfn);
 684         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
 685         unsigned long *subsection_map;
 686         int rc = 0;
 687
 688         subsection_mask_set(map, pfn, nr_pages);
 689
 690         subsection_map = &ms->usage->subsection_map[0];
 691
 692         if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
 693                 rc = -EINVAL;
 694         else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
 695                 rc = -EEXIST;
 696         else
 697                 bitmap_or(subsection_map, map, subsection_map,
 698                                 SUBSECTIONS_PER_SECTION);
 699
 700         return rc;
 701 }
 702 #else
 703 static struct page * __meminit populate_section_memmap(unsigned long pfn,
 704                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
 705                 struct dev_pagemap *pgmap)
 706 {
 707         return kvmalloc_node(array_size(sizeof(struct page),
 708                                         PAGES_PER_SECTION), GFP_KERNEL, nid);
 709 }
 710
 711 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
 712                 struct vmem_altmap *altmap)
 713 {
 714         kvfree(pfn_to_page(pfn));
 715 }
 716
 717 static void free_map_bootmem(struct page *memmap)
 718 {
 719         unsigned long maps_section_nr, removing_section_nr, i;
 720         unsigned long magic, nr_pages;
 721         struct page *page = virt_to_page(memmap);
 722
 723         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 724                 >> PAGE_SHIFT;
 725
 726         for (i = 0; i < nr_pages; i++, page++) {
 727                 magic = page->index;
 728
 729                 BUG_ON(magic == NODE_INFO);
 730
 731                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 732                 removing_section_nr = page_private(page);
 733
 734                 /*
 735                  * When this function is called, the removing section is
 736                  * logical offlined state. This means all pages are isolated
 737                  * from page allocator. If removing section's memmap is placed
 738                  * on the same section, it must not be freed.
 739                  * If it is freed, page allocator may allocate it which will
 740                  * be removed physically soon.
 741                  */
 742                 if (maps_section_nr != removing_section_nr)
 743                         put_page_bootmem(page);
 744         }
 745 }
 746
 747 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
 748 {
 749         return 0;
 750 }
 751
 752 static bool is_subsection_map_empty(struct mem_section *ms)
 753 {
 754         return true;
 755 }
 756
 757 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
 758 {
 759         return 0;
 760 }
 761 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 762
 763 /*
 764  * To deactivate a memory region, there are 3 cases to handle across
 765  * two configurations (SPARSEMEM_VMEMMAP={y,n}):
 766  *
 767  * 1. deactivation of a partial hot-added section (only possible in
 768  *    the SPARSEMEM_VMEMMAP=y case).
 769  *      a) section was present at memory init.
 770  *      b) section was hot-added post memory init.
 771  * 2. deactivation of a complete hot-added section.
 772  * 3. deactivation of a complete section from memory init.
 773  *
 774  * For 1, when subsection_map does not empty we will not be freeing the
 775  * usage map, but still need to free the vmemmap range.
 776  *
 777  * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
 778  */
 779 static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
 780                 struct vmem_altmap *altmap)
 781 {
 782         struct mem_section *ms = __pfn_to_section(pfn);
 783         bool section_is_early = early_section(ms);
 784         struct page *memmap = NULL;
 785         bool empty;
 786
 787         if (clear_subsection_map(pfn, nr_pages))
 788                 return;
 789
 790         empty = is_subsection_map_empty(ms);
 791         if (empty) {
 792                 unsigned long section_nr = pfn_to_section_nr(pfn);
 793
 794                 /*
 795                  * Mark the section invalid so that valid_section()
 796                  * return false. This prevents code from dereferencing
 797                  * ms->usage array.
 798                  */
 799                 ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
 800
 801                 /*
 802                  * When removing an early section, the usage map is kept (as the
 803                  * usage maps of other sections fall into the same page). It
 804                  * will be re-used when re-adding the section - which is then no
 805                  * longer an early section. If the usage map is PageReserved, it
 806                  * was allocated during boot.
 807                  */
 808                 if (!PageReserved(virt_to_page(ms->usage))) {
 809                         kfree_rcu(ms->usage, rcu);
 810                         WRITE_ONCE(ms->usage, NULL);
 811                 }
 812                 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
 813         }
 814
 815         /*
 816          * The memmap of early sections is always fully populated. See
 817          * section_activate() and pfn_valid() .
 818          */
 819         if (!section_is_early)
 820                 depopulate_section_memmap(pfn, nr_pages, altmap);
 821         else if (memmap)
 822                 free_map_bootmem(memmap);
 823
 824         if (empty)
 825                 ms->section_mem_map = (unsigned long)NULL;
 826 }
 827
 828 static struct page * __meminit section_activate(int nid, unsigned long pfn,
 829                 unsigned long nr_pages, struct vmem_altmap *altmap,
 830                 struct dev_pagemap *pgmap)
 831 {
 832         struct mem_section *ms = __pfn_to_section(pfn);
 833         struct mem_section_usage *usage = NULL;
 834         struct page *memmap;
 835         int rc;
 836
 837         if (!ms->usage) {
 838                 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
 839                 if (!usage)
 840                         return ERR_PTR(-ENOMEM);
 841                 ms->usage = usage;
 842         }
 843
 844         rc = fill_subsection_map(pfn, nr_pages);
 845         if (rc) {
 846                 if (usage)
 847                         ms->usage = NULL;
 848                 kfree(usage);
 849                 return ERR_PTR(rc);
 850         }
 851
 852         /*
 853          * The early init code does not consider partially populated
 854          * initial sections, it simply assumes that memory will never be
 855          * referenced.  If we hot-add memory into such a section then we
 856          * do not need to populate the memmap and can simply reuse what
 857          * is already there.
 858          */
 859         if (nr_pages < PAGES_PER_SECTION && early_section(ms))
 860                 return pfn_to_page(pfn);
 861
 862         memmap = populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
 863         if (!memmap) {
 864                 section_deactivate(pfn, nr_pages, altmap);
 865                 return ERR_PTR(-ENOMEM);
 866         }
 867
 868         return memmap;
 869 }
 870
 871 /**
 872  * sparse_add_section - add a memory section, or populate an existing one
 873  * @nid: The node to add section on
 874  * @start_pfn: start pfn of the memory range
 875  * @nr_pages: number of pfns to add in the section
 876  * @altmap: alternate pfns to allocate the memmap backing store
 877  * @pgmap: alternate compound page geometry for devmap mappings
 878  *
 879  * This is only intended for hotplug.
 880  *
 881  * Note that only VMEMMAP supports sub-section aligned hotplug,
 882  * the proper alignment and size are gated by check_pfn_span().
 883  *
 884  *
 885  * Return:
 886  * * 0          - On success.
 887  * * -EEXIST    - Section has been present.
 888  * * -ENOMEM    - Out of memory.
 889  */
 890 int __meminit sparse_add_section(int nid, unsigned long start_pfn,
 891                 unsigned long nr_pages, struct vmem_altmap *altmap,
 892                 struct dev_pagemap *pgmap)
 893 {
 894         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 895         struct mem_section *ms;
 896         struct page *memmap;
 897         int ret;
 898
 899         ret = sparse_index_init(section_nr, nid);
 900         if (ret < 0)
 901                 return ret;
 902
 903         memmap = section_activate(nid, start_pfn, nr_pages, altmap, pgmap);
 904         if (IS_ERR(memmap))
 905                 return PTR_ERR(memmap);
 906
 907         /*
 908          * Poison uninitialized struct pages in order to catch invalid flags
 909          * combinations.
 910          */
 911         if (!altmap || !altmap->inaccessible)
 912                 page_init_poison(memmap, sizeof(struct page) * nr_pages);
 913
 914         ms = __nr_to_section(section_nr);
 915         set_section_nid(section_nr, nid);
 916         __section_mark_present(ms, section_nr);
 917
 918         /* Align memmap to section boundary in the subsection case */
 919         if (section_nr_to_pfn(section_nr) != start_pfn)
 920                 memmap = pfn_to_page(section_nr_to_pfn(section_nr));
 921         sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
 922
 923         return 0;
 924 }
 925
 926 void sparse_remove_section(unsigned long pfn, unsigned long nr_pages,
 927                            struct vmem_altmap *altmap)
 928 {
 929         struct mem_section *ms = __pfn_to_section(pfn);
 930
 931         if (WARN_ON_ONCE(!valid_section(ms)))
 932                 return;
 933
 934         section_deactivate(pfn, nr_pages, altmap);
 935 }
 936 #endif /* CONFIG_MEMORY_HOTPLUG */